license: other license_name: qwen-research-license license_link: LICENSE base_model: Qwen/Qwen-Image-2.1 base_model_relation: quantized pipeline_tag: text-to-image library_name: diffusers tags:
- text-to-image
- image-editing
- diffusers
- nvfp4
- modelopt
- abliterated
- vllm
- dgx-spark
- built-with-qwen
Image-2.1-abliterated-NVFP4
NVFP4 (ModelOpt) checkpoint of an abliterated Qwen/Qwen-Image-2.1, packaged for vLLM-Omni on NVIDIA Blackwell / DGX Spark (GB10).
The behavior edit is a weight-space merge, not a new training run. A rank-128 adapter at step 32,000 was fused into the diffusion transformer at scale 1.0 before quantization. The text encoder and the VAE are the upstream weights.
Research only: refusal-style filtering has been substantially removed. This checkpoint is intended for interpretability, image-model robustness study, red-teaming, and controlled experiments — not for production or end-user deployment.
⚠️ Disclaimer — read before use
This model has had its safety filtering substantially removed by merging an abliterated adapter into the image transformer. As a direct consequence:
- It will follow image prompts that the original Qwen-Image-2.1 would decline or blank out. It has no meaningful built-in guardrails.
- It is released strictly for legitimate research — interpretability, robustness evaluation, red-teaming, and controlled experiments.
- You assume full responsibility and liability for how you use it and for everything it generates. Do not deploy it to end users or in production without adding your own safety, moderation, and abuse-prevention layers.
- Use must comply with the Qwen Research License Agreement inherited from the base model, and all laws and regulations that apply to you. Commercial use needs a separate license from Qwen.
- The authors and uploaders accept no liability for any misuse or harm. Its outputs do not reflect the views of the uploaders or of Qwen / Alibaba.
By downloading or using this model you acknowledge and accept the above.
Provenance
| Step | Detail |
|---|---|
| Base | Qwen/Qwen-Image-2.1 |
| Edit | Step-32000 rank-128 adapter merged at scale 1.0 into the DiT (224 modules). Text encoder and VAE unmodified |
| Quantization | NVFP4 via NVIDIA ModelOpt 0.47.0, algorithm max. Blocks 2–29 are W4A4, group 16. Blocks 0, 1, 30, 31 and the boundary projections (img_in, txt_in, modulation, time_text_embed, norm_out, proj_out) stay BF16. Calibration: 16 prompts, 1024×1024, 16 steps |
| Format | ModelOpt NVFP4 diffusers pipeline. DiT about 5.5 GB; full pipeline about 23 GB |
| Target | vLLM-Omni on Blackwell (sm_121), including DGX Spark / GB10 |
transformer/config.json sets quant_method: modelopt and quant_algo: NVFP4. Do not pass --quantization fp8. This is not a ComfyUI file and not a GGUF. Diffusers from_pretrained does not yet load this NVFP4 schema.
vLLM-Omni (DGX Spark / GB10)
docker run --gpus all --ipc=host -p 8000:8000 \
-e NCCL_CUMEM_ENABLE=0 \
-v "$HOME/.cache/huggingface:/root/.cache/huggingface" \
vllm/vllm-omni:qwen-image21 \
vllm serve GoldfingerCH/Image-2.1-abliterated-NVFP4 --omni --port 8000
Endpoints: POST /v1/images/generations and POST /v1/images/edits.
Send num_inference_steps on every request. The upstream checkpoint wants 40 steps and true_cfg_scale 1.0 when there is no negative prompt. The adapter was evaluated at 25 steps, true_cfg_scale 4.0, and a negative prompt of a single space. Sending a scale above 1 without a negative prompt doubles the compute. The server default is 50 steps.
This checkpoint was built on one DGX Spark. It has not yet been timed under vLLM-Omni on that machine.
License
Qwen Research License Agreement, 20 September 2026, inherited from Qwen/Qwen-Image-2.1. Research and evaluation only. Redistribution must keep LICENSE and Notice. Products built with this model display Built with Qwen.
Credits
- Qwen Team (Alibaba) — Qwen-Image-2.1 architecture and weights
- ChatR Labs — step-32000 adapter
chatrlabs/Qwen-Image-2.1-Uncensored-Omni-LoRA-v0.1 - Merge, NVFP4 quantization, and Spark packaging — GoldfingerCH